Energy Optimization in Wireless Medical SystemsUsing Physiological Behavior

Hyduke Noshadi, Foad Dabiri, Saro Meguerdichian,Miodrag Potkonjak and Majid Sarrafzadeh

Computer Science DepartmentUniversity of California, Los Angeles, California, United States{hyduke, dabiri, saro, miodrag, majid}@cs.ucla.edu

ABSTRACTWearable sensing systems are becoming widely used for avariety of applications, including sports, entertainment, andmilitary. These systems have recently enabled a variety ofmedical monitoring and diagnostic applications in WirelessHealth. The need for multiple sensors and constant monitor-ing lead these systems to be power hungry and expensive, withshort operating lifetimes. In this paper, we introduce a novelmethodology that takes advantage of the influence of humanbehavior on signal properties and reduces those three metricsfrom the data size point of view. This, in turn, directly influ-ences the wireless communication and local processing powerconsumption. We exploit intrinsic space and temporal corre-lations between sensor data while considering both user andsystem behavior. Our goal is to select a small subset of sen-sors to accurately capture and/or predict all possible signalsof a fully instrumented wearable sensing system. Our ap-proach leverages novel modeling, partitioning, and behavioraloptimization, which consists of signal characterization, seg-mentation and time shifting, mutual signal prediction, andsubset sensor selection. We demonstrate the effectiveness ofthe technique on an insole instrumented with 99 pressure sen-sors placed in each shoe, which cover the bottom of the entirefoot, resulting in energy reduction of 56% to 96% for errorrates of 5% to 17.5%.

Categories and Subject DescriptorsC.3 [Special-Purpose and Application-Based Systems]:[Real-time and embedded systems]; B.8.2 [Performance andReliability]: Performance Analysis and Design Aids; J.3 [Lifeand Medical Sciences]: Medical information systems

General TermsDesign, Algorithm, Performance

KeywordsWearable Medical Systems, Energy Optimization, Sensor Se-

Permission to make digital or hard copies of all or part of this work forpersonal or classroom use is granted without fee provided that copies arenot made or distributed for profit or commercial advantage and that copiesbear this notice and the full citation on the first page. To copy otherwise, torepublish, to post on servers or to redistribute to lists, requires prior specificpermission and/or a fee.Wireless Health ’10, October 5–7, 2010, San Diego, USACopyright 2010 ACM 978-1-60558-989-3 ...$10.00.

lection, Behavioral Sensing.

1. INTRODUCTIONEmbedded networked systems and wide area cellular wire-

less systems are becoming ubiquitous in applications rangingfrom environmental monitoring to urban sensing. Meanwhile,sensor networks have emerged as an important class of dis-tributed embedded systems capable of solving a variety ofchallenging monitoring and control problems in a number ofapplication domains, ranging from government and militaryapplications to seismic, habitat, and wildlife continuous ob-servations. These technologies have recently been adopted tosupport the emerging work in medical devices equipped withsensors, known collectively as Wireless Health [1] [2] [3] [4].Wireless Health merges data, knowledge, and wireless com-munication technologies to provide health care and medicalservices such as prevention, diagnosis, and rehabilitation out-side of the traditional medical enterprise.

Such sensor systems have high potential to significantly im-prove the quality of life for large segments of the populationand enable conceptually new types of applications. However,it is important to note that a path to industrial realizationhas been more elusive than initially was expected due to a va-riety of issues, including system and operational complexity,cost and energy sensitivity, semantic complexity, and the needfor often revolutionary changes in consumer behavior. Ever-increasing opportunities in health care have thus motivatedresearchers in Computer Science and Electrical Engineeringto develop technologies that can be adopted in the medicaland physiological fields and to serve the recently growing de-mand of low cost and widely accessible health care services.

In this paper, we show how signal processing techniques(time-shifting and segmentation), in addition to a new com-binatorial optimization paradigm (pseudo-exhaustive combi-natorial search), can be used to design an energy optimizedembedded sensing system to reduce energy consumption bymore than an order of magnitude. While some of these tech-niques best perform on embedded sensing systems that sharelocal communication, a majority of them can be applied onessentially any sensing systems. Our goal is to demonstratethat often expensive wearable sensing systems used in medicalstudies can be made more attractive to daily usage througha system of coordinated design and operational techniquesthat facilitate mass production, customization to specific cus-tomers, and low power operation.

Specifically, our optimization goal is to simultaneously min-imize the cost (i.e. the number of sensors) and energy con-sumption (i.e. the weighted sum of collected and communi-

128

cated samples) while preserving a specified accuracy of col-lected data, or vice versa. To do so, we exploit intrinsic spaceand temporal correlations between sensor data while consid-ering both user and system behavior. Our proposed method-ology takes advantage of signal semantics and predictabilityamong sensors to reduce the number of sensors and amountof data acquisition, and has the following technical novelties:

Signal time-shifting: In many sensing systems, relative timeshifting greatly improves predictability. This phenomenon isstrongly expressed in medical sensing systems. Another im-portant observation is that cross-correlation functions are al-most always unimodular. Therefore, binary search can beused for very fast calculation of the best shifts. In addition,note that the complexity of the sensor selection problem doesnot increase since we can always shift the selected signals byany required amount.

Signal segmentation for mutual sensor prediction: Signals inmany types of embedded sensing systems have natural phases.For example: temperature and humidity are often highly im-pacted by sun activity, which is composed of morning, af-ternoon, and night phases; a heart beat has systolic and dias-tolic phases; and shoe pressure sensors are subject to airborne,landing, and take-off phases. Once the signals are aligned us-ing signal time-shifting, the prediction of signals in each phaseis much more accurate after segmentation, because data in onesegment will otherwise often act as noise for data in another.

Subset node selection: It is easy to see that the selection ofsubsets of sensors from which the values of all other sensorscan be computed within a given error is a NP-complete prob-lem by observing that it can be mapped to the dominating setproblem, which we solve using a novel type of constructive al-gorithm that facilitates an easy trade-off between the qualityof the solution and the run time. Combinatorial iterative com-ponent assembly (CICA) iteratively builds a number of partialsolutions that are likely to be part of the final solution. It canbe easily shown that an arbitrarily close approximation can beachieved at the expense of run time. Much more importantly,CICA has very strong practical performance.

We discuss the related work in the next section, then presentthe low power wearable sensing system that has been the mainmotivation behind this study in Section 3. In Section 4, wepresent the signal properties of this system, which are influ-enced by user behavior and of which we take advantage foroptimization. In Section 5, we mathematically formulate therelationships between pairs of sensors, derive the predictor-to-base sensor model, and define the predictor selection ob-jectives. Finally, Section 7 presents experimental results andour achieved performance.

2. RELATED WORKIn this section, we briefly survey the most directly related

work in (low power) sensor networks, medical and wearablesensing systems, energy optimization in body sensor networks,and sensor reading prediction.

Convergence of sensing, communication, computation, andstorage technologies created the notion, testbeds, theory, andconceptual foundation for sensor networks. The research anddevelopment interest resulted in an exponentially growing num-ber of sensor network publications. There are several concep-tual and comprehensive sensor network surveys [5] [6]. Fromthe very beginning, it was realized that energy is one thestrictest constraints in many classes of sensor networks [7] [8].

With growing interest in designing sensor-based medical de-

vices, wearable embedded sensor systems attracted an inten-sive and fast growing research and industrial interest [1] [9] [10][11] [12] [13]. The emphasis has been on feasibility, processing,and interpretation of medical signals. However, energy opti-mization is especially important in the medical domain, wherelow cost and ease of every day use is crucial. Thus, it has beentargeted by a range of researchers from communication andsignal processing to hardware design and software engineeringin body area networks [14] [15] [16] [17] [18]. There are surpris-ingly few reports related to cost and energy minimization formedical sensing systems. To the best of our knowledge, thereare no reported techniques for simultaneous minimization ofused sensors and the energy budget in wearable systems.

Exploration of the correlation of sensor readings is prob-ably the most addressed task in embedded sensing. Thereare a large number of techniques ranging from a priori as-sumed dependency (e.g. Gaussian and random Markov fields)and similarity to movie streams [19] to non-parametric studiesthat exploit properties of signals such as monotonicity [20].

3. PRELIMINARIESWe will demonstrate our proposed methods for an instance

of expensive systems used in medical studies for daily andubiquitous usage. The target system is a lightweight smartshoe capable of sensing plantar pressure, movement, direc-tion, and rotation. This system can be very attractive fora range of applications, such as instability and gait analysisoutside of a laboratory environment, outdoor gaming, sports,workplace safety, and environmental data collection. In al-most all of these specified applications, long term and contin-uous operation is required, while in an outdoor environmentcharging the batteries of the system is not a convenient oreven possible task. High sampling rate, continuous data col-lection, and large-volume data transmission has introducedtremendous challenges to operating such a mobile platform.Considering the aforementioned issues, it is essential to de-velop a new method for instrumenting the shoe with sensorsin a way that would reduce the total system’s energy con-sumption. In Section 3.1, we explore the architecture of thedesigned lightweight system.

In general mobile or lightweight embedded sensing systems,wireless communication is provably one the most power hun-gry units in the system. In the system under study in this pa-per, we have used MicroLEAP as our main sensor node, whichis responsible for sensing and transmitting the collected datafrom the sensors. Table 1 summarizes the power and energyconsumption of the radio and the processor on MicroLEAP[27].

Table 1: MicroLEAP energy consumption.Power Data Rate Energy/Bit(mW) (kbps) (nJ/bit)

Processor 2.7 NA NARadio 57.5 250 230

Therefore, one can easily conclude that in a network of Nsensors, eliminating the sampling of a subset of the sensor datacan drastically reduce the required energy consumption. Keepin mind that our goal is not to disregard the data from thosesensors but provide a means to retrieve the information lateron. In other words (as described in detail in Section 5), in thisresearch one of our goals is to sample only a small subset of

129

Figure 1: Right: medical shoe instrumented withpressure sensors. Left: pressure sensor location mapunder the foot for Pedar’s insole. Sensor placementand location are derived from studying Pedar’s insole,which has 99 sensors

sensors, and later, in the base station, predict accurately thedata values of the whole sensor network.

3.1 System Set Up and InstrumentationThe designed smart shoe is instrumented with pressure sens-

ing material and an embedded data acquisition unit with pro-cessing and radio transmission capability. For the pressuresensing material, we either use passive resistive sensors pro-duced by Tekscan [28] or the piezoresistive fabric produced byEeonyx [29]. When using passive resistive sensors, sensors areplaced under the insole and are connected to the data acqui-sition unit. In the case of the piezoresistive fabric, the fabricwill be cut such that it covers the entire surface under thefoot, and two conductive layers will be placed on both sidesof the fabric. The sensing areas are the parts of the fabriccovered by conductive material on both sides. In order torecord the pressure values, the sensing area will be connectedto the data acquisition board. The processing unit samplesdata from pressure sensors at 60 Hz. In addition to pressuresensors, the medical shoe also has gyroscopes and accelerome-ters, which are used for activity recognition, motion tracking,and gaming applications.

Sensor placement, especially for resource-constrained sys-tems, requires sufficient understanding of the environmentwhere the sensors are being deployed so that (1) decisive andimportant areas are covered and (2) resource usage is done in-telligently, meaning one does not deploy more resources thannecessary. Therefore, we require a deeper understanding ofthe pressure distribution and behavior beneath the foot sothat we can meet all the required objectives and system con-straints for sensor placement. These objectives can includeenergy requirements, sampling accuracy, coverage, etc.

To better understand the signals resulting from the exertionof pressure by both feet, we use a plantar pressure mappingsystem that covers the entire surface under the foot. Theadvantage of using such a system is that it gives us a clearpicture of the complete pressure distribution under the foot.Pedar [30] is an accurate and reliable pressure distributionmeasuring system for monitoring local loads between the footand the shoe. It is comprised of insoles equipped with a gridof 99 pressure sensors, which cover the entire area under thefoot, and a data acquisition unit capable of data sampling andtransmission to a PC over a wireless (bluetooth) connection.Even though systems such as Pedar can provide complete in-formation about the plantar pressure, high data sampling andtransmission rates make systems such as Pedar unpopular forlow power applications, due to the short lifetime of the system.

Figure 2: Three steps with three extracted stateseach: (A) airborne, (B) take-off, and (C) landing.

4. SIGNAL PROPERTIESWe study plantar pressure signal properties corresponding

to human ambulation in order to identify physiological andbehavioral trends. The extracted patterns and signal seman-tics, along with behavioral properties, are used in this studyto model the relationship among plantar pressure signals.

A plantar pressure signal can be segmented based on its be-havior, which is imposed by human gait characteristics. Wedivide each step into three segments: (1) airborne; (2) land-ing; and (3) take-off. Figure 2 demonstrates the extractedsegments in the plantar pressure signal. The airborne stateis defined as the time during which a particular foot is nottouching the ground. The landing state is defined as the timefrom when the signal starts increasing its amplitude from thebase offset value (calibrated zero pressure) until exactly beforeit starts decreasing its value; the landing state for each partof the foot, then, is the time during which the body’s weightis applied to that particular sensor. Finally, the take-offs tateis the time interval during which the signal’s amplitude de-creases from its peak to the base offset value.

Pressure signal characteristics such as morphology, ampli-tude, and pattern are influenced by an individual’s physiologyand walking behavior. For example, Figure 4 shows pressurereadings from all 99 sensors recorded from Pedar for two testsubjects, two steps each, where one had flat feet and the otherhad hollow feet. As the figure suggests, the active pressurearea is greater for the flat-footed person, while the amplitudedifference between the active pressure and passive pressure ar-eas for the hollow-footed person is much higher. Furthermore,the pressure patterns and signal morphology are almost thesame in each step for the same test subject though differentbetween the two. It is important to note that the morphologi-cal similarity will not be guaranteed if the test subject changesthe type of ambulation (e.g. walking, jumping, or standing),even though it is consistent within the same type of activity.

Furthermore, the maximum amplitude of a signal is depen-dent on the relative position of the sensor to the person’s cen-ter line of pressure progression under the feet; sensors closerto the center line of pressure will record higher pressure val-ues compared to those that are located at the border of theactive and passive pressure areas. Sensors that are locatedon the center line of pressure or on the lines parallel to itdemonstrate almost identical behavior but at different times.Therefore, we can divide the sensors in the active pressure area

130

0 50 100 150 200 250 300 350 400 4500

50

100

150

200

250Non Shifted

Time elapsed

Ampl

itude

0 50 100 150 200 250 300 350 400 4500

50

100

150

200

250Shifted

Time elapsed

Ampl

itude

Figure 3: Top: base and predictor signals at origi-nal times. Bottom: base signal time-shifted towardpredictor signal, where resulting signals are almostidentical.

into sets, where data extracted from all sensors in the sameset have similar morphology and almost identical shape whenshifted. This implies that the signal’s behavior is propagatingin the walking direction onto different sensors in the same set.We take advantage of consistent progression in data modelingand predictor selection. Figure 3 shows two signals at theiroriginal times and when one is shifted toward the other.

5. PREDICTIONMODELINGIn order to predict the behavior of the sensors from each

other, it is essential to use a good prediction model to mini-mize prediction errors. Due to sensors’ diverse locations andtheir behavior under the foot during human motion, it is im-possible to have a single fitting model to be used as the predic-tion function across sensor pairs. Therefore in order to avoidcost and complexity of managing many prediction models, wetake advantage of shifting signals. Due to consistent prop-agation of applied pressure under the active pressure sensingarea, there exists a shift for a potential base sensor toward thepredictor sensor’s direction, which will align two sensor valuessuch that they will have an overlap between their landing andtake-off states. Our measurements show that once a base sen-sor is shifted toward a predictor sensor such that there is anoverlap among their landing and take off states, we will need 3different mathematical models to present the best predictionfunction between any pair of sensors. The first fitting model isa linear function, while the other two are isotonic. The linearmodel is the best predictor when two sensors have completeoverlap between landing and take-off states. The other twoisotonic models, which are composed of piecewise linear andquadratic models, are the best predictors when take-off andlanding states of the base and predictor sensors are not com-pletely aligned together, and either or both are aligned withthe other’s airborne state.

Figure 5 demonstrates the fitted linear curve, and the iso-tonic curves, which illustrate the mathematical relationshipbetween values from two different sensors, namely the predic-tor and base sensor.

5.1 Prediction ErrorWe have considered two objectives for prediction accuracy

while fitting the data using the above specified prediction func-tions. The first objective was to create the model such that it

minimizes the sum of the squares of the residuals as describedin Equation 1, which is basically the least-squared method(ls). The second objective was to minimize the sum of theabsolute values of the residuals as described in Equation 2,otherwise known as the L1-model.

minimize(sqrtm∑

t=1

r(t)2) (1)

minimize(m∑

t=1

|r(t)|) (2)

We evaluate the whole process of sensor predictor selectionusing both of these error definitions.

5.2 Predictor Selection ObjectivesOur proposed methodology in this section is aimed at select-

ing a potentially small subset of deployed sensors along withprediction functions such that by only utilizing that smallset of sensors, all sensing data can either be measured di-rectly or predicted with an acceptable error bound. Considertwo sensors si and sj and assume the corresponding sensorvalues as functions of time are denoted as gi(t) and gj(t).For every pair of sensors we create a collection of predictorsΦij = {φij1, ...,φ ijm}. φijk represents a predictor function forsensor sj which is based on shifted values from sensor si by ksamples. In other words, if the predicted value for sensor sjis denoted by g∗j (t) we have:

g∗j (t) = φijk(gi(t− k)) (3)

For a given predictor there is a prediction error associatedwith it. We use a different cost function for prediction erroras described in Section 5.1. For instance, least square basedprediction error can be presented as:

ε(φijk) =

∑Tt=1

(gj(t)− φijk(gi(t− k)))2

gj(t)2(4)

For a given sensing system, the prediction transform matrixis defined as below:

Ψl×n =

. . .φijk

. . .

(5)

where l ≤ n is the number of predictors denoted by P ={p1, ..., pl} ⊆ S. Now we can formally define the problem.The sensor predictor selection objective can be formulated as:

minimize(l = |P |) (6)

such that:

∀1 ≤ j ≤ n, ∃i, s.t.Ψij &= ∅ (7)

∀Ψij &= ∅, ε(Ψij) ≤ δ (8)

Constraint 7 guarantees that for any given sensor there is atleast one predictor, whereas constraint 8 enforces that max-imum prediction error for any given sensor is less than thetarget threshold of δ, where δ is an input to the problem de-fined by the user or is application driven.

131

0

20

40

60

80

100

020

4060

80100

1200

200

400

Sensor Id

Flat foot

Samples

Ampl

itude

(a)

0

20

40

60

80

100

0

20

40

60

80

1000

100

200

300

Sensor Id

Non Flat foot

Samples

Ampl

itude

(b)

Figure 4: Pressure mapping under the feet for (a) flat feet and (b) hollow feet. The progression of pressuresensors over the active pressure area is observable in both cases. The locations of sensors under the foot arebased on the pressure sensor map in Figure 1.

0 50 100 150 200 2500

50

100

150

200

250

Predictor Sensor Value

Base

Sen

sor V

alue

(a)

0 20 40 60 80 100 120 140 160 180 200

0

50

100

150

200

Predictor sensor value

Base

sen

sor v

alue

(b)

0 20 40 60 80 100 120 140 160 180 200

0

50

100

150

200

Predictor sensor value

Base

sen

sor v

alue

(c)

Figure 5: Relationship between predictor vs. base sensor when: (a) the take-off and landing states are overlap-ping; (b) the landing or take-off state of the base is overlapping with the airborne state of the predictor; and(c) the landing or take off state of the predictor is overlapping with the airborne state of the base.

6. OPTIMUM PREDICTOR SELECTIONIn this section we cover the steps involved in the sensor

selection process.The first step of the process is to generate the prediction

functions φijk. Each sensor is potentially a predictor. In thecontext of prediction, we refer to predictor sensors as pi andthe sensors being predicted as base sensors, si. For a givenpredictor sensor pi we generate n × m predictor functions:{φijk} where 1 ≤ j ≤ n and 1 ≤ k ≤ m. A prediction errorcorresponds to each predictor function represented as εijk,which is computed using Equation 4. Predictor functions aregenerated as described in Section 5.1.

The top predictor set of sensor sj is defined as:

Tj = {pi1 , ..., pik}s.t.εpijk ≤ δ (9)

For each sensor that is a top predictor (i.e. ∀pi ∈ ∪Tj), wecreate a set of base sensors for which that predictor is amongthe top predictors. In other words:

πpi = {sj1 , ..., sjl}, s.t.∀sj ∈ πpi , pi ∈ Tsi (10)

Basically, πpi represents the sensors which can be predictedby sensor i with prediction error less than δ.

6.1 Combinatorial Iterative Component Assem-bly

The goal in sensor selection is to find a minimal set of pre-dictors, Π= {pi}, which can be used to predict all othersensors. Formally this objective can be stated as minimizing|Π| such that:

∀sj ,∃pi ∈ Π, s.t.pi ∈ Tj (11)

We call this minimal setΠ ∗. The way we tackle this problemis to select a minimum number of πpis which cover the wholeset of sensors. This problem is equivalent to the minimum setcover problem which is known to be NP-Hard. Therefore, weuse a combinatorial iterative component assembly algorithm,or CICA, to find the min set cover. We compare the perfor-mance of CICA with a well known approximation algorithmdescribed in [32]. In the experimental results, we show thatCICA is indeed performing better than the aforementionedapproximation. CICA works in the following way. First itwill sort the set of predictors based on the maximum numberof sensors they can cover. Then it picks the top predictorsfrom the sorted list and combines each with the initial list tocreate a new predictor-to-base sensor map. This process con-

132

Algorithm 1 Minimum Set Cover Using Combinatorial Iter-ative Component Assembly

1: Input: πpi for all sensors in the system and k for top setselection threshold

2: Output:Π ∗ minimum set of sensors, which can be usedto predict other sensors

3:Γ ← Sort πpis based on number of sensors they can predictin descending order and pick top K sets

4: Υ= { }5: Index = { }6: while no set in Γ covers all sensors in the system do7: for each set γi inΓ do8: for each πpi from Input do9: Combine covered sensors in γi and πpi and add the

new set toΥ10: Add predictor sensor to γi’s corresponding index.11: end for12: end for13:Γ ← Sort sets in Υ based on number of sensors they

cover in descending order and pick top k sets14: end while15:Π ∗ ← Index corresponding to largest set inΓ16: returnΠ ∗

tinues until there is at least one single set which covers all thesensors. Algorithm 1 summarizes the process.

OnceΠ ∗ is created, we generate the Φ matrix, and the sen-sor selection process is completed. The way we create theelements in the Φ matrix is as follows. Rows of the matrixcorrespond to the predictors inΠ ∗ and the entries of the ma-trix are:

Φ[pi, j] = argmin(ε(Ψpijk)), ifpi ∈ Tj (12)

Φ[pi, j] = ∅, otherwise (13)

In other words, in the column corresponding to base sensorsi, we insert the best prediction function from its top predic-tors. Algorithm 2 summarizes the process.

Algorithm 2 Minimum Predictor Selection

1: Create prediction functions, φijk

2: Find top predictors Tsi for every base sensor si3: Using Tjs, create the set of base sensors (πpi) best pre-

dicted by every top predictor pi.4: Find the minimum set cover from πpis and add the corre-

sponding predictors toΠ ∗

5: UseΠ ∗ to create the prediction matrixΦ

6.2 Generalized Sensor SelectionIn general, sensor networks are deployed in a particular envi-

ronment to collect specific information from that environment.Any optimization of configuration methodology will be appliedto that sensor network either prior to deployment or after-ward. Many of the environment-dependent offline method-ologies need to be repeated if a new sensor network is to bedeployed in a new place. For the case study in this paper, thesensing environment is the human body (in particular, thefeet). Therefore, for any new test subject, some training datashould be collected to efficiently select the best predictors andcustomize the system accordingly. At the same time, it is only

1 2 3 4 5 6 7 8 9 100

20

40

60

80

100

120

140

Prediction error vs number of selected sensors forincremental set of subjects

Number of Subjects

Pred

ictio

n er

ror

2.5% error5% error7.5% error10% error

12.5% error15% error17.5% error

Figure 6: The number of predictor sensors convergesfor the generalized case.

natural to assume that, across different people, the sensing en-vironments have some similarities that might render repeat-ing predictor selection for any new test subject redundant. Toovercome this problem, we also tried to find the globally min-imum top predictors across all subjects. To achieve this, wesimply modified the process as follows. We defined πpi to be:

πpi = {sj1O1, ..., sjlOr

} (14)

where Ok represents the kth subject, and r is the total num-ber of subjects under study. Basically, base sensors are dif-ferentiated across subjects with the secondary Ok index, butpredictors remain the same. Therefore, the number of basesensors to be covered by min set cover is increased by a factorof the number of test subjects. The rest of the process remainsthe same.

The main question to address is: how stable are the globalpredictors? In other words, if the top predictors are selectedbased on training data from k test subjects, how will thosepredictors perform for the (k+1)th test subject? Experimen-tally we show that once the number of test subjects for globalpredictors is around 5, the corresponding predictors are in factglobal and reliable for any new subject. Figure 6 shows thenumber of predictors for various numbers of test subjects anderror rate bounds. We generated this graph by running, for agiven number of test subjects (say k), the generalized predic-tor selection of all combinations of k test subjects for whomwe had reported the average predictor size. It is clear fromthis graph that the predictor size converges very quickly oncethe number of test subjects passes 7. This means that, oncethe global predictor set is calculated for a few test subjects,these predictors can be reliably used for a new subject.

7. EXPERIMENTAL RESULTSIn order to illustrate the effectiveness of the proposed method-

ologies, we used Pedar to collect sensor data across 10 indi-vidual subjects. We tried to keep the subject set as diverseas possible in terms of walking behavior and sensor data vari-ability. The subject set was composed of 7 men and 3 women,where one man and one woman were flat-footed and 2 menwere overweight. Foot sizes ranged from 7 to 11.

For the collected data set, we performed minimum predictorselection with three different configurations: 1) least-squaremethod for predictor function generation without any shiftin base-sensor data (ls-noShift); 2) least-square method for

133

2.5 5 7.5 10 12.5 15 17.520

30

40

50

60

70

80

90

100

Average prediction error rate (%)

Num

ber o

f pre

dict

ors

ls−no shiftls−with shiftls−with shift & segmentationL1

(a)

2.5 5 7.5 10 12.5 15 17.50

10

20

30

40

50

60

70

80

90

Average prediction error rate (%)

Num

ber o

f pre

dict

ors

ls−no shiftls−with shiftls−with shift & segmentationL1

(b)

Figure 7: Number of predictor sensors vs. the average prediction error for the whole sensing network for (a)the general case and (b) averaged over individual test subjects.

2.5 5 7.5 10 12.5 15 17.50.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Average prediction error rate (%)

Ener

gy s

avin

g ra

tio

ls−no shiftls−with shiftls−with shift & segmentationL1

(a)

2.5 5 7.5 10 12.5 15 17.50

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

Average prediction error rate (%)

Ener

gy s

avin

g ra

tio

ls−no shiftls−with shiftls−with shift & segmentationL1

(b)

Figure 8: Ratio of energy savings with respect to no prediction vs. the average prediction error for the wholesensing network, for (a) the individual case and (b) the general case.

predictor function generation with shift in base sensor value(ls-wShift); 3) segmentation-based predictor function genera-tion using least-square method (ls-segmentation); and 4) L1-method for predictor function generation (L1).

Furthermore, we repeated the above scenarios for the globalcase where the process was implemented on the aggregateddata from all individuals (as described in Section 6.2). 30% ofeach test subject’s data was used to find the best predictor,and the remaining 70% was used to evaluate the accuracy ofthe prediction functions. We ran the sensor selection processfor a range of maximum prediction errors (δ in Equation 8).The maximum prediction error rate ranges from 2.5% to 20%.Afterwards, we simulated the estimated total energy savingson the sensor node (MicroLEAP) when the minimal predic-tors were used to sample the data. As seen in the systemarchitecture, one sensor node is responsible for sampling andtransmission of the data from sensors.

Table 2 summarizes the performance of the proposed com-binatorial iterative component assembly algorithm to find theminimum set cover versus the greedy approach. As the ta-ble suggests, our proposed algorithm (CICA) outperforms thewell known greedy algorithm in both the individual and gen-eral case. In the individual case, CICA outperforms the greedyalgorithm by an average of 22.3% for error rates between 2.5%and 12.5%, while for 15% and 17.5% errors, the outcome is al-most the same for both algorithms. In the general case, CICA

outperforms the greedy algorithm by an average of 23.8% forerror rates greater then 5%, while for 2.5% the outcome ofboth algorithms are the same.

Table 2: Minimum selected sensors using CICA vsGreedy algorithm

Individual GeneralError CICA Greedy Error CICA Greedy2.5 63 78 2.5 94 945 41 52 5 65 827.5 21 27 7.5 59 7510 14 18 10 42 5812.5 11 15 12.5 42 5615 5 6 15 24 3217.5 4 4 17.5 21 27

Figure 7(b) illustrates the number of sensors required topredict the whole sensor network data (out of 99 sensors total)versus the maximum prediction error for the four differentscenarios described above. These graphs are averaged over the10 test subjects we had. Figure 8(a) summarizes the averagepower savings ratio for different maximum prediction errors.These graphs illustrate that power saving ranges from 18% to91%, depending on the approach taken, while the maximumprediction error rate ranges from 2.5% to 20%.

Finally, we repeated the same set of experiments and simu-

134

lations for the general case. Figures 7(a) and 8(b) summarizethe results. As expected, the general case tries to find a globalpredictor whose size is usually larger than the predictor setof an individual, and therefore energy saving ranges from 0%to 76%. Note that for a maximum prediction error rate of2.5%, we must practically select all the sensors, resulting inno energy savings.

8. CONCLUSIONIn this paper, we explored data volume management and

its corresponding implications on energy consumption andsystem lifetime in embedded and wearable sensing systemsthrough the introduction of multiple stages of signal analy-sis and optimization algorithms. The high level contributionof this paper is to study signal patterns and utilize them todevelop novel prediction algorithms at different levels of in-formation flow in such systems. These methods aim to makeexpensive wearable sensing systems more feasible for everydayuse by minimizing sampling sources while enabling reconstruc-tion of the data from all sensors. One goal was to use a subsetof sensors to accurately generate the data from all sensors.This goal was achieved by introducing two novel methods forsignal shifting, which enables better prediction of sensor data,followed by data segmentation to further enable piecewise pre-dictions. To solve the above problems, we also developed anefficient approximation algorithm called combinatorial itera-tive component assembly (CICA) to select optimum predic-tors for each scenario. In order to show the effectiveness ofthe proposed methodologies, we applied the presented meth-ods on an embedded wearable sensing system equipped with100 pressure sensors. Experimental results show that the pro-posed techniques can yield from 56% to 96% in energy reduc-tion while maximum sampling error rate ranges from only 5%to 17.5%.

9. REFERENCES[1] S. C. Jacobsen, T. J. Petelenz, and S. C. Peterson,

“Wireless health monitoring system,” US Patent6,160,478, 2000.

[2] R. Jafari, F. Dabiri, and M. Sarrafzadeh, “CustoMed: Apower optimized customizable and mobile medicalmonitoring and analysis system,”ACM HCI Challengesin Health Assessment, 2005.

[3] W. Wu, A. Baui, M. Batalin, L. Au, J. Binney, and W.Kaiser, “MEDIC: Medical embedded device forindividualized care,”Artificial Intelligence in Medicine,vol. 42, no. 2, pp. 137-152, 2008.

[4] D. Malan, T. Fulford-Jones, M. Welsh, and S. Moulton,“CodeBlue: an ad hoc sensor network infrastructure foremergency medical care,”WAMES, 2004.

[5] D. Culler, D. Estrin, M. Srivastava, and Guest Editors,“Introduction: overview of sensor networks,”Computer,vol. 37, no. 8, pp. 41-49, 2004.

[6] I. F. Akyildiz, T. Melodia, and K. R. Chowdhury, “Asurvey on wireless multimedia sensor networks,”Computer Networks, vol. 51, no. 4, pp. 921-960, 2007.

[7] D. Ganesan, B. Krishnamachari, A. Woo, D. Culler, D.Estrin, and S. Wicker, “Complex behavior at scale: anexperimental study of low-power wireless sensornetworks,” Technical Report UCLA/TR 02-0013, 2002.

[8] J. Polastre, J. Hill, and D. Culler, “Versatile low powermedia access for wireless sensor networks,” Sensys, 2004.

[9] S. Popovic, M. R. Dietz, V. Morari, M. Pappas, I. Keller,and T. Mangold, “A reliable gyroscope-based gait-phasedetection sensor embedded in a shoe insole,” IEEESensors Journal, vol. 4, no. 2, pp. 268-274, 2004.

[10] W. Wu, L. Au, B. Jordan, T. Stathopoulus, M. Batalin,W. Kaiser, A. Vahdatpour, and M. Sarrafzadeh, “TheSmartCane system: an assistive device for geriatrics,”Body Area Networks, 2008.

[11] K. Oshima, Y. Ishida, S. Konomi, N. Thepvilojanapong,and Y. Tobe, “A human probe for measuring walkability,”Sensys, pp. 353-354, 2009.

[12] V. Erickson, A. U. Kamthe, and A. E. Cerpa,“Measuring foot pronation using RFID sensor networks,”Sensys, pp. 325-326, 2009.

[13] K. Lorincz, B. Chen, G. W. Challen, A. R. Chowdhury,S. Patel, P. Bonato, and M. Welsh, “Mercury: a wearablesensor network platform for high-fidelity motionanalysis,” Sensys, pp. 183-196, 2009.

[14] V. Leonov, P. Fiorini, S. Sedky, T. Torfs, and C. VanHoof, “Thermoelectric mems generators as a powersupply for a body area network,” vol. 1, pp. 291-294,2005.

[15] L. Yan, L. Zhong, and N. K. Jha, “Energy comparisonand optimization of wireless body-area networktechnologies,” BodyNets, pp. 1-8, 2007.

[16] H. Ghasemzadeh, E. Guenterberg, K. Gilani, and R.Jafari, “Action coverage formulation for poweroptimization in body sensor networks,”ASPDAC, pp.446-451, 2008.

[17] Y. Liu, B. Veeravalli, and S. Viswanathan,“Critical-path based low-energy scheduling algorithms forbody area network systems,” RTCSA, pp. 301-308, 2007.

[18] S. Xiao, A. Dhamdhere, V. Sivaraman, and A. Burdett,“Transmission Power Control in Body Area SensorNetworks for Healthcare Monitoring,” IEEECommunications Journal, vol. 27, no. 1, pp. 37-48, 2009.

[19] S. Goel and T. Imielinski, “Prediction-based monitoringin sensor networks: taking lessons from MPEG,”SIGCOMM, vol. 31 no. 5, 2001.

[20] F. Koushanfar, N. Taft, and M. Potkonjak, “Sleepingcoordination for comprehensive sensing using isotonicregression and domatic partitions,” INFOCOM, 2006.

[21] M. A. Batalin, M. Rahimi, Y. Yu, D. Liu, A. Kansal, G.S. Sukhatme, W. J. Kaiser, M. Hansen, G. J. Pottie, M.Srivastava, and D. Estrin, “Call and response:experiments in sampling the environment,” Sensys, pp.25-38, 2004.

[22] G. Gandhi, S. Suri, and E. Welzl, “Catching elephantswith mice: sparse sampling for monitoring sensornetworks,” Sensys, pp. 261-274, 2007.

[23] M. Malinowski, M. Moskwa, M. Feldmeier, M.Laibowitz, and J. A. Paradiso, “CargoNet: a low-costmicropower sensor node exploiting quasi-passive wakeupfor adaptive asynchronous monitoring of exceptionalevents,” Sensys, pp. 145-159, 2007.

[24] S. Jevtic, M. Kotowsky, R. P. Dick, P. A. Dinda, and C.Dowding, “Lucid dreaming: reliable analog eventdetection for energy-constrained applications,” IPSN, pp.350-359, 2007.

[25] P. Wan and M. D. Lemmon, “Event-triggereddistributed optimization in sensor networks,” IPSN, pp.49-60, 2009.

135

[26] J. Liu, P. Cheung, F. Zhao, and L. Guibas, “Adual-space approach to tracking and sensor managementin wireless sensor networks,”WSNA, 2002.

[27] L. K. Au, W. H. Wu, M. A. Batalin, D. H. Mclntire,and W. J. Kaiser, “MicroLEAP: energy-aware wirelesssensor platform for biomedical sensing applications,”BIOCAS, pp. 158-162, 2007.

[28] Tekscan pressure sensor. http://www.tekscan.com.

[29] Eeonyx resistive fabric. http://www.eeonyx.com.[30] Pedar: plantar pressure system. http://www.novel.de.[31] L. Klingbeil and T. Wark, “A wireless sensor network

for real-time indoor localisation and motion monitoring,”IPSN, pp. 39-50, 2008.

[32] P. Slavik, “A tight analysis of the greedy algorithm forset cover,” STOC, pp. 435-441, 1996.

136